Mistral previewed Large 4, a 1 trillion parameter model, with open weights due October
Mistral AI opened a preview API for Mistral Large 4, a 1 trillion parameter multimodal model, and says open weights will follow by the end of October.
Mistral Large 4 is Mistral's largest model so far, up from the 675 billion parameters of Mistral Large 3. It is a mixture-of-experts model, so each token goes through a small set of expert subnetworks and only 49 billion parameters are active at a time. Those are the figures in Mistral's blog. Mistral's documentation lists 1.05 trillion total and 52 billion active, and the company hasn't explained the gap.
Mistral says it trained the model from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European datacenters. It claims Large 4 significantly outperforms any open-weight model from the US or Europe and is competitive with the strongest open models globally. No independent benchmark results are available yet. For now the preview runs only through the API on Mistral Studio, and Mistral is red-teaming a version with reduced moderation with cybersecurity partners until the weights ship.
On 5 October, Reflection AI unveiled Beam, its first model. Beam is a text-only mixture-of-experts model with 501 billion total and 23 billion active parameters, a 1 million token context window and 23.8 trillion pretraining tokens, and it is aimed at coding and agent tasks. Reflection says Beam matches Z.ai's GLM-5.2 on reasoning benchmarks while using 3 to 4 times less inference compute, and nobody has checked that claim independently. GLM-5.2 is about 744 billion total and 40 billion active parameters. Reflection says the weights and a technical report will come later in October.
Google DeepMind released EmbeddingGemma 2 with open weights on 6 October. An embedding model turns inputs into vectors so that similar things sit close together. This version puts text, code, images, video and audio into one shared 768-dimensional space, where the first EmbeddingGemma handled text only. The encoders are modular, so a developer can load the 270 million parameter text part alone or go up to 740 million with vision and audio. Google reports about 191MB of active RAM for the text-only weights and about 567MB for the full model on a Pixel 11 Pro.
Also in the news
- Google will limit free Gemini users to Flash Lite from 9 October, and the $4.99 per month AI Plus plan will lose Gemini Pro, which will stay on the $19.99 AI Pro and $99.99 Ultra plans.
- Google DeepMind released Nano Banana 2.1, an image generation and editing model priced on OpenRouter at $1.50 per million input tokens, $7.50 per million output tokens and $30 per million image output tokens.
- Vals AI reports that a team of Claude Opus 5.5 agents found two candidate room-temperature antiferromagnetic semiconductors for computer memory, predicted by calculation and not yet tested in a lab.
- Meta published the Personal Agent Protocol with Walmart, Stripe, Sierra and others, an open standard meant to help websites tell legitimate user agents from malicious bots.
- OpenAI announced labeled image ads that appear next to image generation results for ChatGPT Free and Go users, with US testing starting later in October.
- OpenAI reports that GPT-6 Astra scored 55.0% on 11 contracting tasks built with Ironclad, against 41.6% for GPT-5.6 Sol, and the page carries no publication date.
- Anthropic expanded its Cyber Verification Program to absorb Project Glasswing, with three access tiers for vetted cyberdefenders, six days after Google gave Gemini 4 Argon to cyber defenders first.
- Technology Innovation Institute announced Falcon-Emirati-7B, a model built on Falcon-H1-Arabic and specialised in Emirati Arabic dialect and culture.
- DeepSeek is reportedly close to a $12 billion funding round backed by Tencent.
- Kuaishou has reportedly picked banks for a Hong Kong IPO of its Kling video unit worth more than $1 billion, according to The Information.
- OpenAI will reportedly start watermarking ChatGPT text in the EU.